Crucible
A mixture-of-experts model compression toolkit that prunes, merges and quantizes large MoE language models into smaller specialist models for local inference, implementing REAP expert pruning with several scoring strategies and a self-contained evaluation harness.
Crucible compresses mixture-of-experts language models by removing experts that contribute little to a target specialist role: because a MoE model activates only a handful of its experts per token, for example 8 of 128, most of the resident weight for any single role goes largely unused. Crucible scores each expert by how much work it actually does on calibration data representative of that role, using the REAP pruning method with several scoring strategies, then removes the low-scoring experts and leaves a model that quantizes and serves normally afterward.
The project is architecture-agnostic: supporting a new MoE model family only requires adding a single ModelAttrs registry entry rather than a code or method change. It ships as a uv workspace holding two independent packages - crucible, which prunes, merges and quantizes to produce weights, and crucible-bench, a separate evaluation harness that drives any OpenAI-compatible endpoint to score a model and can be installed alone without the heavier PyTorch dependency.
It requires Python 3.12 or later and a transformers release new enough to support recent Gemma and Qwen MoE architectures; quantized output uses the compressed-tensors format vLLM serves natively.